The v2.2.0 version bump (2.1.27 -> 2.2.0) updated the crate Cargo.toml files
but not Cargo.lock, leaving the lockfile pinning vestige-core/vestige-mcp at
2.1.27. The Release workflow builds with `cargo build --locked`, which refuses
to update the lockfile, so every platform target failed with:
error: cannot update the lock file ... because --locked was passed
Regenerated via `cargo update -w` (workspace members only, no dependency bumps).
Verified: `cargo build --locked -p vestige-mcp --release` now succeeds.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Opens with the pitch's pattern-interrupt first sentence ('Your bug was born
days before it crashed'), then tells the story in Sam's own voice — why he built
it, the soccer/causal-gap framing, the DeepMind theorem + CauseBench receipts
(0% vs 60%, kept honestly separate as theorem-vs-measurement). Links the
60-second spoken pitch (demo/PITCH-v2-causebench.md). Real citations only
(arXiv:2508.21038, Nature DOI 10.1038/s41586-024-08168-4).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Complete rewrite that leads with the visceral problem (the agent that forgets
the lesson you taught it), then the v2.2.0 flagship (Retroactive Salience
Backfill — memory with hindsight) and the claim-contradiction breakthrough.
Updates 25→13 tools, v2.1.23→v2.2.0, refreshes all metrics (86k LOC, 23MB,
1550 tests). Real citations only (Zaki/Cai 2024 Nature 637:145-155, no
fabricated URLs). All internal doc links verified to exist.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Second swarm pass (complete every-line sweep), verified against real code
(38/101 confirmed real; 63 false positives excluded). This commit lands the
main-compatible subset:
- redmine SSRF guard: rewrite host check to use host_str()+std::net::IpAddr
instead of the `url` crate (url is not a direct dep of vestige-core on main;
the previous form only compiled on the feature branch). Same protection:
blocks localhost + loopback/private/link-local/unspecified IPs.
- bin/restore: guard wrapper[0] index (empty backup array would panic)
- bin/cli: char-boundary-safe node.id truncation (2 byte-slice panics);
XML-escape the model/home strings before launchd plist substitution
- prospective_memory: Duration::try_hours/try_days (panic on out-of-range
user config); case-insensitive " at " split now uses the lowercased index
so the contains() check and the split agree
core 477/0, mcp builds, clippy clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Verified main-compatible moderate fixes from the complete sweep:
- predictive_retrieval get_proactive_suggestions: clone session_context and
drop the read guard before predict_needed_memories re-acquires it (re-entrant
RwLock read can deadlock when a writer is queued between)
- hippocampal_index create_semantic_associations: snapshot (id, summary) pairs
under the read lock, drop it, THEN run the O(n) cosine scan (was blocking all
writers for the full scan duration)
- prospective_memory check_triggers: don't auto-expire an intention that was
JUST triggered this iteration (the expire ran unconditionally after
mark_triggered, clobbering a fresh trigger)
core 477/0, clippy clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Tool Consolidation v2.2.0 (34->12 advertised surface: recall/memory/codebase/
intention/smart_ingest/source_sync/memory_status/maintain/dedup/graph/
session_start/suppress) onto the retroactive-salience-backfill base.
Conflict resolution:
- server.rs: took #99's canonical 12-tool surface, then re-added
as a 13th ADVERTISED tool (ToolDescription + dispatch arm + test). Backfill
is the flagship 'Memory with hindsight' (Cai 2024 Nature) cognitive primitive
— a distinct capability, not a maintenance op to fold into `maintain` or hide.
- Cargo.toml: kept usearch features=["fp16lib"] with the fuller MSVC-C1021
explanation comment (Windows build fix, #71).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Wire Step 8.5 into run_consolidation (bounded + idempotent via durable
Causal edge), share looks_like_failure/extract_entities across MCP tool +
CLI + auto-fire, add whole-word failure markers (API_TIMEOUT no longer
false-fires), --json CLI for CauseBench, and refactor the Redmine SSRF
guard to std::net (drops the dangling url-crate dep).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Adds docs/launch/tool-consolidation-v2.2.0.md — the single sequenced plan that
reconciles the two prior planning notes:
- Layer 1 (this PR): 34 → 12 advertised tools, safe commit order, alias policy,
preserved invariants, and the test that proves each.
- Layer 2 (follow-up): tiny always-on default surface + SessionStart/Stop hooks.
Also refreshes stale in-code comments to match the consolidated surface:
- server.rs handle_tools_list header (was "v2.1.21: 25 tools") and the
size-annotation rationale (now lists recall/memory_status/dedup/graph).
- tools/mod.rs module doc (the facade vs. granular-handler relationship).
No behavior change. Gates: cargo test --workspace, cargo clippy -D warnings,
pnpm dashboard check + build — all green.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Tool Consolidation v2.2.0, Layer 1 commit 5/6. Advertised tools 21 → 15.
Folds consolidate + dream + gc + importance_score + backup + export + restore
into one action-dispatched tool:
action = consolidate | dream | gc | importance_score | backup | export | restore
- Thin facade forwards the same args envelope to each existing handler; the
`action` discriminator is cloned out before the envelope is moved into a
callee. No handler uses deny_unknown_fields, so per-action params validate.
- Safety defaults preserved (all handler-internal): gc dry_run=true by default,
restore path-confinement, export traversal guard. Verified by
test_maintain_actions_and_safety (gc dry-run + restore missing-path error).
- Events preserved end-to-end:
- Pre-dispatch Started events (ConsolidationStarted/DreamStarted) re-emitted
in the `maintain` arm keyed on action.
- emit_tool_event normalizes the `maintain` name to its effective action, so
the existing ConsolidationCompleted/DreamCompleted/ImportanceScored arms
fire unchanged — no duplicated emit logic.
- All 7 old names remain hidden warn!+redirect aliases (removed v2.3.0), keeping
their own pre-emits.
- Tests: count 21→15, 7 negatives, new dispatch/safety test.
Gates: cargo test --workspace, cargo clippy -D warnings — clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Tool Consolidation v2.2.0, Layer 1 commit 4/6 (3b). Advertised tools 24 → 21.
Folds explore_connections + predict + memory_graph + composed_graph into one
action-dispatched tool:
action ∈ {chain, associations, bridges, predict, memory_graph,
recent, get, memory, neighbors, never_composed, bounty_mode, label}
- Transparent facade: each action forwards the same args envelope to the
existing handler, which re-reads its own discriminator/params. No underlying
arg struct uses deny_unknown_fields, so cross-fields are ignored.
- All actions read-only except `label` (the one mutator), which is logged for
audit via tracing::info!(event_id, outcome_type).
- outcome_type enum sourced from composed_graph::OUTCOME_TYPES (now pub(crate))
rather than re-listed, so the vocabulary stays single-sourced.
- All 4 old names remain hidden warn!+redirect aliases (removed v2.3.0).
- Size annotation: graph=250_000 (memory_graph layout + bounty_mode pagination),
kept in sync across loop, helper, and both annotation tests.
- Tests: count 24→21, 4 negatives, test_graph_actions_and_aliases exercising
read-only actions + aliases (incl. the memory_graph facade branch).
Note: the design draft mislabeled graph::execute as sync; it is `pub async fn`.
The per-commit build gate caught it — the facade awaits it correctly.
Gates: cargo test --workspace, cargo clippy -D warnings — clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Tool Consolidation v2.2.0, Layer 1 commit 3/6 (3a). Advertised tools 27 → 24.
Folds system_status + memory_health + memory_timeline + memory_changelog
into one view-dispatched tool:
view = health (default) | retention | timeline | changelog
- Thin facade: each view forwards the same args envelope to the existing
handler. No underlying arg struct uses deny_unknown_fields, so the `view`
discriminator is ignored by each handler and per-view fields validate as
before. The cognitive lock is never held across a forwarded call.
- view='health' returns the byte-for-byte system_status shape (audit scripts
parse it), incl. schema_introspection passthrough — verified by
test_default_view_is_health asserting equality with execute_system_status.
- All 4 old names remain hidden warn!+redirect aliases (removed v2.3.0).
- Size annotation moved: memory_timeline (200k) → memory_status, kept in sync
across the real loop, expected_max_result_size(), and both annotation tests.
- Tests: count 27→24, 4 negative asserts, test_memory_status_views_and_aliases
exercising all 4 views + 4 aliases.
Gates: cargo test --workspace, cargo clippy -D warnings — clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Tool Consolidation v2.2.0, Layer 1 commit 2/6. Advertised count unchanged
at 27 (pure rename).
`session_start` is the imperative-outcome name for the one-call session
initializer. `session_context` remains a hidden warn!+redirect alias
(≥1 minor release, removed v2.3.0), calling the same handler unchanged.
Tests: positive assert swapped to session_start + negative assert for the
old name. Gates: cargo test --workspace, cargo clippy -D warnings — clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
All verified against real code:
- github fetch_updated: propagate fetch_comments errors instead of
unwrap_or_default() (silent failure stored a comment-less record with a
corrupted hash AND advanced the cursor past it = permanent gap)
- relationships load_relationships: advance next_id past loaded rel-N ids so
new_id() can't collide with persisted ids
- speculative store_pending_predictions: merge instead of clear (two predict()
calls without an intervening record_usage() wiped the first batch's accounting)
- embeddings/code is_comment_only: stop treating leading '#'/'*' as comments
(was deleting #include, #define, *ptr, CSS '*', multiplication)
- importance_signals: reconcile total_count to the surviving pattern sum after
decay/prune (was inflated, driving novelty toward zero over time)
- importance on_retrieved_with_context: set context in the SAME critical section
as the event push via record_retrieval() (two-lock approach raced; last_mut()
could land on a concurrently-pushed event)
core 535/0, clippy clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
All verified against real code:
- connectors: empty list_live_ids() no longer mass-tombstones the entire source
(treat empty as "cannot enumerate", like None) — was catastrophic data loss
- redmine: SSRF guard — require http(s), reject loopback/private/link-local hosts
+ localhost (escape hatch VESTIGE_ALLOW_PRIVATE_CONNECTOR_HOSTS for local tests)
- KnowledgeEdge::is_valid now honors valid_from via was_valid_at(now) (was
ignoring it; a future-dated edge read as valid). No production callers.
- emotional_memory: flashbulb counter reconciles to the FINAL is_flashbulb
decision after the importance override (was undercounting via ==0 guard)
- chains path_to_chain: fixed off-by-one — step i now uses the INCOMING edge
connections[i-1], not the outgoing connections[i]; removed the synthetic
connection hack that masked the misalignment
- merge_supersede compose_merged_content: exact normalized dedup instead of
substring containment (was silently dropping "cat" inside "cathedral")
core 535/0, clippy clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Multi-model audit (deepseek-v4/minimax/kimi/qwen) surfaced these; verified
against the real code and fixed the confirmed ones:
- github connector: host-pin failed OPEN when api_root was unparseable/hostless
— the bearer token would ride a Link `next` url to an attacker host. Now
fail-closed: no pinned host => drop the url. (CRITICAL: SSRF / token exfil)
- GithubConfig/RedmineConfig derived Debug leaked the token/api_key into any
{:?} log line or panic message. Replaced with manual redacting Debug impls.
- cross_project priority calc used `as u32 - i` which underflows/panics (debug)
or wraps + corrupts the sort (release). Use saturating_sub.
Verified false-positive (no change): path-traversal in get_file_context — it
only inspects the path string, never reads the file.
core: 535 passed / 0 failed, clippy clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Sam's locked stage pitch for a multi-pitch event + separate networking hour.
Keeps the World Cup soccer analogy (the weapon that makes anyone get "causally
upstream"), the detonation line + silence, "realizes", and "next layer of the
dev stack". Closer fits reality: "come find me after, I'll show you in 30s" (demo
is 1:1 at networking on a pre-loaded laptop, never installing on their system).
Two seams closed: "hours"->"days", "agents that touch" (plural). Added networking
playbook: pre-warm the seeded demo, demo from strength, the pilot-hook answer for
"does it work on mine?".
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Sam's canonical 60s spoken pitch, every sentence workshopped. Two seams closed
from the assembled draft: opener "hours"->"days" (matches "three days ago"), and
the reveal names "Vestige" (was a naked "memory"). Includes the win-the-room
line list, stage rules (never throat-clear the open; dare don't demo; end on
"come prove me wrong"), and a delivery map of where to slow/pause/drop voice.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Sam chose to keep Vestige over Postdict: it sounds natural, it's the live + loved
product (561 stars, repo, npm packages), costs zero to keep, and "vestige = a
trace of what remains from the past" is backward-looking = on-thesis. The
category lives in the tagline ("the first memory that finds the cause, not the
lookalike"), not the name. Reverted the pitch to "I created Vestige from my tiny
Chicago apartment", kept the "our"->"my" solo-founder fix.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Sam's own tightened version, with 2 corrections: "our memory" -> "my memory"
(solo-founder first-person rule) and the name is Postdict (not Vestige). Adds
the "tiny Chicago apartment" line (underdog-relatable, investor-loved). The 2024
Nature reference verified true. This is the canonical stage script.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Sam is a solo founder and built Postdict alone — the pitch must say "I built,"
"I ported," "I already did it," never "we." For an early-stage solo founder that
is a strength (ships without a team = what investors bet on). Only remaining
"we" is "now we're handing debugging to AI agents" — there it means all of us as
an industry, not the company. Added delivery rule #4: "I built this" is a flex.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Sam delivers this on a stage to 50-100 people — voice only, no screen. Rewrote
from a terminal-demo script into a pure ~60s spoken pitch built for the ear:
short sentences, hard stops, [pause] cues (silence = the loudest move on stage),
and one memorizable detonation line ("a root cause never looks like the bug it
creates"). Opens about the audience's pain, not the product; promises the
laptop demo afterward rather than fumbling a terminal live. Includes a 30s
hallway version and a 5s one-liner.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The detonation version of the funding pitch. Deliberately REJECTS a "X companies
got hacked" hook (commodity security-FUD, needs an uncitable stat that torches
the honest+reproducible moat, and shrinks the market from "every agent in
production" to "security"). Instead opens with the WOUND every engineer has
lived — "production breaks, the cause is a change you made days ago and forgot"
— then detonates the industry's flawed axiom ("relevance ≠ resemblance; a root
cause never looks like the bug it creates"), proves it live, defends the moat
(Nature port + their architecture IS the axiom), and ends on TAM.
Uses `postdict` on screen (the new name); includes the interactive-shell alias
so it can be recorded with the new name today. Contrast re-verified end-to-end.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Rewrote the 60s funding script around Sam's stronger framing: lead with the
flawed AXIOM the whole industry shares ("relevance equals resemblance"), then
detonate it with the fact ("a root cause never looks like the bug it creates").
This names the MECHANISM of the error, not just that one exists — a diagnosis,
not a claim. Keeps the live contrast as the proof, the Nature-port + "their
architecture IS the axiom" as the moat, and "every agent that touches
production" as the market. On-screen commands re-verified end-to-end:
similarity→billing lookalike (wrong), Postdict→API_TIMEOUT cause (right).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A separate, investor-grade script (vs the viral clip): leads with the thesis
"the entire AI-memory industry is trapped in a category error — a root cause
never looks like the bug it creates", proves it live with the similarity-vs-
Postdict contrast, then closes on the moat (faithful Nature port + incumbents'
architecture IS the category error) and the market (every agent that touches
production).
Verified the exact on-screen commands end-to-end: the 3-memory scenario (cause
+ billing-500 lookalike + crash) makes the contrast devastating — similarity
search returns the billing lookalike as its #1, Postdict reaches back 3 days to
the real env-var cause. (Without the lookalike distractor the contrast collapses
— similarity would also surface the cause — so the script plants all three.)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
demo/README.md: the complete self-serve demo artifact — one-command run, the
seeded scenario explained, a "build your own scenario" section, the honest
boundary (won't invent a cause; can't reach a cause that was never recorded),
the Nature citation + the "field admits this is unsolved" sources, and the
recording playbook + paste-ready caption.
Writing/testing the README surfaced a real inconsistency, now fixed:
- The CLI's failure-finder used a hardcoded content-only marker subset and
ignored tags, so a "Checkout latency spiked" memory (regression tag, no crash
word in content) was never picked as the failure. The CLI now calls the SAME
public `looks_like_failure` (content + tags, full list) the backfill tool uses
— one definition, no drift.
- Extended FAILURE_MARKERS with performance/degradation failures (spiked,
latency, degraded, slow, hang, throttled, oom, 502/503/504, flaky, ...) so the
feature backfills from perf regressions, not just hard crashes.
clippy clean; 527 core + 453 mcp tests; both the main demo and the README's
custom scenario verified end-to-end.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
3-model rotation audit (DeepSeek V4-Pro / Kimi K2.7 / MiniMax M3, max thinking,
each model × each of 3 sections). Claude verified every finding against code.
CONFIRMED + FIXED:
- [FTS, consensus DeepSeek+MiniMax] sanitize_fts5_or_query split on
!is_alphanumeric()+'_', but the index uses tokenize='porter ascii' which
splits on '_' and non-ASCII. So "API_TIMEOUT"/"café" became single phrases that
could NEVER match. Now splits on !is_ascii_alphanumeric() + lowercases to mirror
the tokenizer; caps token count (64) and length (64) for DoS hardening. Also
fixes the pre-existing storage.search bug (multi-word queries silently returned
nothing). 5 new tests pin it.
- [Demo honesty, consensus Kimi+DeepSeek] the contrast labeled keyword search as
"SIMILARITY SEARCH" and asserted "NONE of these is the cause" universally. Now
prints the REAL engine ("keyword (BM25)" vs "semantic (vector + BM25 hybrid)")
and claims only what's true ("ranked by RESEMBLANCE; its top hit is a lookalike").
De-hardcoded the "Service crashed:" munging to a generic label-strip.
VERIFIED FALSE POSITIVE (not changed): MiniMax "fts.id non-existent column" —
the FTS5 table is declared `fts5(id, content, tags, ...)`, the JOIN is valid.
No injection found by any model (quote-doubling + operator-stripping confirmed safe).
clippy clean; 527 core + 453 mcp tests pass; demo verified.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
CLI surface for Retroactive Salience Backfill so the seeded demo (and anyone who
clones it) can reproduce "memory with hindsight" from a terminal:
- `ingest --ago-days N`: backdate a memory N days (plant a dated cause/history).
- `backfill [--failure-id ID] [--manual] [--lookback-days N] [--no-promote]`:
reach backward from a failure and surface+promote the causal earlier memory,
with demo-grade colored output (↩ reached back N days, 🔗 causal join: <entity>,
similarity rank, ✅ promoted).
Verified live end-to-end on a real DB: plant a 3-day-old API_TIMEOUT env-var note
+ a semantically-similar 500 distractor + a crash, run `vestige backfill`, and it
surfaces the env-var note by the shared api_timeout entity (ignoring the similar
distractor) and promotes it. clippy clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The MCP surface for memory-with-hindsight. When a failure memory exists, the
`backfill` tool reaches backward across the real store and promotes the quiet
earlier cause that a vector search structurally cannot surface (not similar to
the failure, only causally upstream via a shared entity).
- tools/backfill.rs: builds BackfillCandidates from real KnowledgeNodes (entities
from tags + heuristic env-var/path/identifier extraction), computes real cosine
similarity from stored embeddings (to PROVE the cause ranks low on similarity),
runs the core RetroactiveBackfill, and promotes surfaced causes via
storage.promote_memory. Auto-finds the latest failure, or takes failure_id;
manual=true forces; promote=false for a dry run.
- registered + dispatched in server.rs (35 tools now); tool-list test updated.
- storage: added pub set_created_at (backdate created_at) so the demo/test can
plant a dated cause.
LIVE RECEIPT: live_backfill_surfaces_root_cause_through_storage ingests a
3-day-old API_TIMEOUT env-var note + a semantically-similar 500-error distractor
+ a crash into a REAL SQLite store, runs the backfill tool, and asserts it
surfaces + promotes the env-var note by the shared API_TIMEOUT entity (the root
cause RAG misses). clippy clean; 522 core + 453 mcp tests pass.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The headliner neuro-mechanism: when a salient FAILURE lands (bug/crash/
regression — the "aversive event"), reach BACKWARD in time and promote the
quiet earlier memory that caused it — the one a vector search structurally
cannot surface because it isn't *similar* to the failure, only causally upstream.
Faithful port of Zaki/Cai et al. 2024, Nature 637:145-155 ("Offline ensemble
co-reactivation links memories across days"), causally proven (hippocampal
silencing abolishes the linking). Ported faithfully:
- backward-only asymmetry (fear links retrospectively, never prospectively) —
also exactly correct for software: a root cause is always upstream in time.
- linking flows along the shared-entity overlap (same file/env-var/service),
NOT semantic similarity — that's the whole point (RAG already covers similarity).
- scoped to failure->backward-causal-backfill, not "all salience flows backward"
(mirrors the Cai aversive->neutral paradigm; honest about scope).
Trigger: auto-detect (high prediction-error + failure markers) OR manual override.
Promotion: boosts FSRS stability so the cause stops decaying and surfaces next time.
Receipt (4/4 tests): backfill_surfaces_the_cause_rag_misses proves it promotes a
sim=0.11 env-var note over a sim=0.82 distractor by the shared API_TIMEOUT entity;
backward-only (future memory never promoted); no shared entity => no fabricated
cause; non-salient doesn't fire; manual override works. clippy clean; 522 core
tests pass (no regressions).
Wires into existing primitives: prediction_error gate (salience), dreams/
consolidation (offline window), memory/strength (promotion). MCP tool + live
demo next.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Lets registries such as Glama start the server in an isolated sandbox and
run the standard MCP stdio introspection (tools/list, resources/list,
prompts/list) so the listing passes its checks and scores.
- node:20-slim (glibc) base, required because the npm postinstall downloads
the prebuilt x86_64-unknown-linux-gnu binary (a -gnu binary will not run on
musl/Alpine).
- Installs vestige-mcp-server@latest globally; postinstall fetches the binary.
- VESTIGE_DATA_DIR=/data keeps memory state inside the container.
- ENTRYPOINT vestige-mcp speaks MCP over stdio.
Verified locally: npm install + postinstall fetch the binary, the server boots,
runs its migrations, and answers initialize + tools/list over stdio.
Two deeper review findings (both blockers) + doc de-staling.
C2-deep: my earlier C2 made purge/delete TRACE as memory.write, but gate_writes
did `get_node(id) -> skip on None`, and purge had already DELETEd the row — so a
destructive removal still never opened a Memory PR (it was silently skipped).
The most security-critical write type couldn't be reviewed. Fix: a missing node
is now gateable for destructive decisions — gate_writes builds the WriteContext
from the decision itself (marks `forgets`, which classify_write gates), and the
PR records the removal with node.deleted=true. Proven live: purging a node opens
a PR (kind node_decayed, deleted true); test
gate_opens_pr_for_destructive_write_after_node_deleted_c2.
PRIV: gate_writes copied the FULL node.content into the PR diff + title, so a
real secret in a gated memory would leak into the memory_prs table, the
dashboard, and any exported proof bundle — defeating the point of gating
sensitive writes. Fix: the PR now stores a truncated content PREVIEW + an FNV
content HASH, and sensitive-topic/sensitive-node-type writes are fully REDACTED
("[redacted — sensitive content; review via risk signals]"). The reviewer still
sees the risk signals (why it opened) and a hash (to correlate), never the
secret. Tests gate_redacts_sensitive_content_in_pr_priv,
content_preview_redacts_sensitive_and_truncates, content_hash_is_stable. The
committed memory_pr.json + the whole proof bundle were re-captured and contain
no secret (verified by scan); the re-shot memory-prs.png shows the redaction.
DOC: REVIEW.md commit list is now git-log-based (no stale hashes); C2-deep + PRIV
added to the findings table; PROOF.md write/PR rows updated; test count -> 1007.
Gates: 1007 lib tests pass (+7 new regressions), clippy -D warnings clean,
dashboard check + build clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Two more review findings — both real, both blockers — plus stale-doc cleanup.
C1: the B1 release used reverse_suppression(subject_id, labile_hours), which
REFUSES once the 24h active-forgetting labile window has passed. So a Memory PR
reviewed late could be marked "promoted" while its memory stayed suppressed.
Approving a quarantined write is an explicit reviewer decision and must release
the memory regardless of elapsed time. New SqliteMemoryStore::release_quarantine
fully clears the suppression (count→0, suppressed_at→NULL) with NO time-window
limit; the PR handler now uses it. Proven: a test backdates suppressed_at to
+100h, shows reverse_suppression refuses, and release_quarantine still releases.
C2: memory(action="purge"|"delete") returns `action` + nodeId but those labels
weren't in is_write_decision, so destructive removal bypassed the memory.write
trace and the PR gate. Added purge/purged/delete/deleted/forget/forgotten.
Proven live: purging a node now records a second memory.write event
({"decision":"purge"}) under the run.
Docs: REVIEW.md de-staled — removed the fixed 140b15f diff-stat / "3 commits"
prose (it moved with each fix), listed all commits, added C1/C2 to the findings
table, updated the test count.
Gates: 1002 lib tests pass (+3 new regressions), clippy -D warnings clean,
dashboard check + build clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A full multi-agent review found 7 real issues (4 blockers). All fixed + tested.
B1 (blocker): Promoting a Memory PR did not release the quarantined memory —
the UI said "promoted" while the memory stayed suppressed/out of retrieval.
act_on_memory_pr now calls reverse_suppression(subject_id) on accept actions;
MemoryPrAction::releases_memory() encodes the rule (promote/merge/supersede
release; forget/quarantine keep it held). Proven live: PR response
subjectReleased:true, SQLite suppression_count 0.
B2 (blocker): memory promote/demote (returns `action`, not `decision`) and
codebase remember_* writes bypassed the write-trace + PR gate. extract_writes
now reads `action` too, filtered by is_write_decision (reads like get/state
excluded); is_write_tool includes `codebase`.
B3 (blocker): receipt ids collided within a run (r_<date>_<runId> +
INSERT OR REPLACE overwrote earlier receipts). IDs are now
r_<date>_<runId8>_<unique6>; build() mints the suffix, build_with_unique()
keeps tests deterministic.
B4 (blocker): proof bundle was assembled from two runs (trace.json=run_proof,
websocket-events.jsonl=run_proof2). Re-captured the whole bundle from a single
run — trace, websocket, receipt, and memory_pr all carry run_proof now.
B5: Black Box receipts panel showed global latest, not the selected run.
Added list_receipts_for_run + /api/receipts?run= ; the page uses listForRun.
B6: SENSITIVE_TOPICS substring matching false-fired (tokenizer->token,
author->auth, secretary->secret). Switched to word-boundary matching; real
phrasings (auth token, security vulnerability, api key) still gate.
B7: set_review_mode now writes atomically (temp+rename via write_atomic);
export_trace sanitizes run_id in the Content-Disposition filename; memory-prs
static routes declared before the dynamic /{id} route.
Withdrawn: the /mode-vs-/{id} route order is NOT a functional bug (axum 0.8 /
matchit prioritizes static segments) — reordered for clarity only.
Gates: 999 lib tests pass (+9 new regressions), clippy -D warnings clean,
dashboard check + build clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Freeze feature work and package the Black Box proof bundle for full review
before anything else lands (no quarantine-constellation work has started).
REVIEW.md is the single entry point: frozen public claim, the per-producer
caveat ledger (real / caveat — sanhedrin.veto stays an honest off-by-default
caveat; dream.patch is REAL but not live-by-default), and the exact command
receipts (run live at 2026-06-22T22:57:59Z, toolchain pinned) for every
validation status credited — nothing claimed from memory.
The bundle (blackbox-proof-2026-06-22/) already contains: status/trace/receipt/
memory_pr JSON, two live WebSocket captures (one with the dream.patch
TraceEvent), six screenshots (Black Box, Receipts/ReceiptCard, Memory PRs,
Graph, dream Black Box, dream producers panel), and PROOF.md. Every credited
artifact was verified to contain real content.
Review surface: 27 source files, +5830/-18 across commits 80c823a, b89beee,
140b15f (build artifacts + proof bundle excluded).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Bounded follow-up (tight acceptance criteria, no scope expansion): flip the
dream.patch producer from "quiet because no dream ran" to a recorded live event.
The dream tool's `insights` array carries no per-item id, so the recorder
extracted zero proposals and dream.patch never fired even on a real dream.
Fix: derive a stable proposal id from each insight's REAL content (its
insight_type + the source memories it consolidated). The dream genuinely ran;
this just gives each real proposal a deterministic handle. Unit-tested against
the exact dream output shape.
Proven end to end (run_dream_proof, 6 memories consolidated):
- one dream.patch event: dream:RecurringPattern:5d941c7f+a41aca72+...
- SQLite + /api/traces/:runId: dream-trace.json (14 events, last is dream.patch)
- WebSocket: dream-websocket-events.jsonl (the dream.patch TraceEvent)
- dashboard: screenshots/dream-producers.png — the row flips to "fired this run"
PROOF.md updated: dream.patch moves from CAVEAT to REAL (still not live by
default — it fires only when a dream actually runs, and the UI says so).
sanhedrin.veto remains an honest CAVEAT (optional hook, off by default).
Gates: 957 lib tests pass, clippy -D warnings clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Make the receipt chain impossible to doubt. Freeze the claim surface, prove
every hop, and turn the two off-by-default producers into explicit UI states.
Frozen public claim: "Vestige records real MCP memory activity into a
replayable local trace, with receipts and reviewable risky writes." We do NOT
claim Sanhedrin vetoes or dream patches are live by default.
Regression — full-spine test (server.rs): one runId must cross, byte-identical,
MCP output -> SQLite trace -> WebSocket event -> API response shape ->
MCP resource. Fails if any hop drops or rewrites the id.
Honest UI states (Black Box "Event producers" panel):
- sanhedrin.veto -> "No veto producer connected (optional Sanhedrin hook, off
by default)" instead of empty mystery.
- dream.patch -> "No dream run in this trace" unless a dream actually ran.
- contradiction.detected -> "no contradiction in this run" when none fired.
Quarantine review (not pre-write blocking): risky writes are committed then
suppressed — audit history preserved, retrieval influence suspended until
reviewed. Reworded the server notice + UI copy to say exactly that.
Receipts UI gap closed: ReceiptCard is now mounted on the Black Box page
(retrieved/suppressed/trust-floor, activation path, "Open receipt in Cinema").
Proof pack (blackbox-proof-2026-06-22/): status.json, trace.json (the
.vestige-trace.json export), receipt.json, memory_pr.json (promoted via
UI->API->SQLite), websocket-events.jsonl (live TraceEvent x6 + PR opened/
decided), screenshots (Black Box, Receipts, Memory PRs, Graph), and PROOF.md
with real/caveat/stub per feature.
Gates: 988 lib tests pass, clippy -D warnings clean, dashboard check + build
clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The loop fired requestAnimationFrame immediately and never awaited render(); with
150k particles + compute + bloom a frame can exceed 16ms, so rAF callbacks queued
faster than the GPU could drain them and the camera/parallax visibly lagged behind
the cursor. Now AWAIT sb.render(dt) before scheduling the next frame → the loop is
capped to real GPU throughput, every frame reflects the latest pointer position,
no backlog. Also snappier active-steer damping (lam 3.5→9, ~110ms converge) so
input feels immediate; idle glide-home unchanged. renderFailures resets on success.
Plus docs/MEMORY_CINEMA.md — complete feature reference for the cinema engine.
Gate: svelte-check 0/0, 937 tests, verified live (parallax tracks cursor, no lag).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
near-fade × fog × DOF × seam-fade each multiply a <1 factor; together they were
crushing the figure dark. Raised the fog floor (0.18→0.45) and the color/emissive
glow bases so the stacked attenuation lands in a vivid range. Verified live.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Step 5 — VELOCITY-STRETCH FLYTHROUGH: sandbox derives camera velocity per frame
(one Vector3, zero compute) and pushes view-space apparent velocity to the storm;
flythrough relaxes the camera clamp floor (lerp 30→6) so the camera plunges
inside the shell. Storm stretches each sprite along screen-space velocity via
rotationNode + scaleNode (clamped streak), separate output graph (no extra
positionView read). Defaults 0 → no-op until wired.
Step 6 — INTERACTIVE PARALLAX: pointer orbits / scroll + pinch zoom the camera
with frame-rate-independent damping, composed onto the director's base pose in
loop() (after director.update, before render); idle >2.5s eases back to 0 so it's
a toy when touched and a film when left alone. sandbox.render re-clamps so the
user can't break framing. Per-beat flythrough strength wired from shot.tension;
dream mode flies through at 0.6. Fully gated off under reduced-motion (no
listeners, flythrough 0).
The 4-feature immersion stack (infinite zoom + flythrough + parallax + DOF/fog)
now composes. Gate: svelte-check 0/0, 937 tests, build green, verified live (all
4 compose, no white-out, no recursion, parallax responds without breaking framing).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The cloud now dives inward FOREVER, seamlessly. Two layers ride offset phases of
fract(uTime/T): the outer grows pow(λ, phase) toward the camera then snaps back
(invisible — inner@1/λ == outer@1); the inner (half-period offset) grows promoted
by λ to become the next outer shell, while a fresh inner spawns inside. λ=1.923
(=1/0.52 inner scale) makes the snap mathematically exact. A sin(phase·π) seam
cross-fade in rimFactor makes each layer fully transparent at its snap → ZERO pop.
Particle-space (not a camera dolly) so it can't clip or fight the camera clamp.
Rack-focus tracks the descent. uZoomOn gates it: Act II+ and dream mode dive;
beats 0/1 + reduced-motion stay still. Verified: seamless loop, no white-out, no
recursion, 937 tests pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Off-focus particles dim (read as bokeh defocus under the bloom) with a breathing
rack-focus. Folded into the single depthFade depth read — NOT sprite scaleNode,
which collapsed the sprites to invisible and collides with the upcoming streak.
Subtle (0.3) so it adds cinematic depth without darkening the figure.
Gate: svelte-check 0/0, 937 tests, verified live (depth grading reads, no recursion).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Distant particles dim toward the void with view depth (exp falloff) → real 3D
atmospheric recession, not a flat sprite cloud. Combined with the near-fade into
a SINGLE depthFade Fn — critical three@0.172 TSL constraint discovered: reading
positionView from a second Fn feeding the same material output triggers a cyclic
stack-overflow in the node type-resolver (getNodeType). One depth read, one Fn.
Gate: svelte-check 0/0, 937 tests, verified live (fog depth reads, no recursion).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
